CalciumFM — how each model sounds (and breaks)

The quality reference: highest fidelity, no knobs — control is only (mean latent, length) morphing + exact inversion.

The 22M deployment model (50 ms/drum on 8 CPU threads); very close to v1_4 in sound.

Length knob commanded −2σ vs +2σ at default guidance: the two clips are nearly identical (the style vector outvotes the knob — pooled ρ≈0.06). Same +2σ with boosted per-knob guidance: now it lengthens.

Invert the kick's noise, then command +2σ attack spread: the edit does nothing (leaky free dims + noise carry everything). Identity is perfect but the model is uneditable this way.

The same invert-and-edit on the same kick: the spread edit lands (single hit becomes a wider burst) while the sound stays recognizably the same kick — the adversarial disentanglement at work.

Four free samples from perturbed library CLAP embeddings — no real sample conditioned, KAD 0.63 (conditioned band ≈0.3). Then the kick with length −2σ vs +2σ: nearly identical — the frozen CLAP outvotes the knobs. v4 samples; it does not edit.

Same free-sampling recipe (CLAP + anchor knobs, KAD ≈0.97) AND the same kick length edit now responds — the scrubbed CLAP projection no longer outvotes the knobs. The one model that both samples and edits; recon is its weak point.

The same kick, length −2σ / 0 / +2σ, same noise — first on v3.1 (hear it drift toward rumble as it lengthens), then on v5 (it stays the same kick, just longer/shorter). Below: v5's new F0 knob at −1.5σ / 0 / +1.5σ — the kick retunes. v5's trade: slightly less polished overall texture (KAD 0.52 vs 0.32).

v6 (released) = v5 + 40k decoder-only polish steps with control pinned by self-distillation from the v5 teacher; encoder frozen, so v6 codes are drop-in with v5's. Grid: 12 sources x the full editing repertoire. Sources marked * were picked from the top of the measured v6-v5 recon-delta distribution — the median per-sample gain is +0.01 cos (subtle on short, easy drums; both models are near the ceiling there), the audible gains concentrate in hard samples like these. † recon at CFG 2.0 + z0x0.7, SAME noise for v5 and v6. Edits run at CFG 2.0 with full-temperature z0 (shrunken noise kills structural knobs — if a knob feels dead in your editor, check the z0 temperature first). PITCH NOTE: a 45 Hz kick sits at the dataset's pitch FLOOR (encoded f0 anchor z=-1.4), so f0 -1.5sigma barely moves it (45->43 Hz) while +1.5sigma retunes it clearly (45->100 Hz) — pitch-down works on mid/high-pitched sources (snare, tom, hat), pitch-up works everywhere. Length edits are RELATIVE, +-0.75 sigma around the encoded length of the source (CFG 1.5): moderate, natural envelope changes (bigger jumps push into synth-like tails). The last column is the recommended identity-preserving structural edit: invert the source, blend 50/50 with fresh noise, regenerate with the edited code.

8 rank-16 adapters over frozen v6, one per CLAP timbre cluster, control pinned by self-distillation (adapters on HF). The win is texture realism: routed recon-KAD 0.186 → 0.135 at the tuned point (real-data floor 0.097); median latent-cos gain is small (+0.001–0.003), so these five sources are picked from the TOP of the measured per-sample delta (+0.02–0.05) — where the difference is audible. Same noise for both recons; editing knobs verified intact (len ρ 1.0, invert-edit 0.97).

v8 makes the INVERSION pathway editable: the VST encodes a sample, inverts it once (identity 0.994 at CFG 1.5 — hear the recon column), then edits knobs directly at that noise. Before v8, a length edit at full identity destroyed the sound (v6 column: kick attack drift -3.7 oct, 8/24 styles dead). v8 was fine-tuned on inversion couplings (pair-reflow): same edit now lands cleanly (+0.1 oct, 0/24 dead, aggregate drift 0.26 — better than any prior model at random noise). Last column: pitch +1.5σ at full identity. All edits are RELATIVE, in z-score units of the library distribution: f0 sigma = 1.64 octaves, so +-1.5 sigma retunes ~+-2.5 octaves from the source pitch (clamped to the +-2.5 sigma training range); scatter = transient count (rolls/claps vs one hit), spread = attack smear.

v9.1 (experimental, not released): the pitch knob retrained WITHOUT any DSP-processed audio — 81,177 natural pairs mined from sample-pack note families (same instrument, different note). Realized pitch span 2.26 -> 3.63 octaves vs v6 (rho 1.0). Rows: f0 -1.5σ / encoded / +1.5σ, same noise, v6 vs v9.1. Length knob currently needs curated pairs (dead on most styles) — next calibration step.